Spatial-frequency domain-aware network for hyperspectral multiclass change detection with subpixel guidance

Hyperspectral image change detection (HSI-CD) is critical for applications such as environmental monitoring, land-use management, and disaster response, where accurate detection of changes over time is essential. Despite recent advances in deep learning-based methods, which leverage the ability to learn complex, nonlinear features from hyperspectral data, several challenges remain. These include the mixed pixel problem, where multiple land cover types within a single pixel make accurate detection difficult, and the issue of multiscale change targets, where changes occur at different spatial scales, complicating the detection of fine-grained changes. To address these issues, this paper introduces SFDANet, a spatial-frequency domain-aware network designed for multiclass change detection in hyperspectral images with subpixel guidance. The main innovations of SFDANet are: (1) subpixel representation learning via spectral unmixing, which allows the model to identify and differentiate change classes at a finer subpixel level; (2) joint spatial and frequency domain feature learning through Fourier convolution, which enables the network to effectively capture both local spatial details and global contextual information, thereby improving detection accuracy; and (3) the adoption of weighted cross-entropy loss to alleviate class imbalance and enhance model robustness. Experimental results validate the effectiveness of joint spatial-frequency feature learning in multiclass change detection tasks, demonstrating substantial improvements in detection accuracy.